In previous posts I have occasionally mentioned an AI product called Warp that I have used successfully in vibe coding and for other AI tasks. Some of the things that Warp can do have greatly impressed me. I have not received any considerations or discounts to use Warp, I subscribed to it and paid for it because I found it to be very useful.
When I first subscribed to Warp, I felt that I was getting more than my money’s worth from it. But over the course of a few months, a combination of changes made by Warp and changes in my usage patterns soured me on Warp. It no longer offers me the value it once did, and I have stopped using it.
Pricing Model
Sometimes tech companies get off to an impressive start, then discover that they are not making as much money as they might like, so they change things to increase their profits. I’m afraid that this may have happened with Warp.
Warp completely changed its pricing model. The new model definitely made Warp more expensive for my (relatively light) usage pattern. Warp switched to a model where a subscription buys some number (e.g. 1500) of ‘AI credits’ each month, and each time you ask it to do something it subtracts from your AI credit balance. But the credits expire at the end of each month, when they charge you the full subscription price and reset your credits to e.g. 1500. Pricing models like this that include a ‘use it or lose it’ aspect (also common in cell phone plans) have always annoyed me. I feel pressured to use credits that I’ve paid for so that they don’t expire and evaporate. The Warp pricing plan has an added twist where you could buy a batch of additional AI credits for more money beyond your subscription price, and those additional credits do not expire each month, they roll over so that they can be used in subsequent months. But frankly this just annoys me more - why not just allow the credits I bought with the subscription to roll over, instead of charging more money for a different class of credits that do roll over? It just makes for a more complicated pricing model that further emphasizes the ‘use it or lose it’ aspect, and it seems designed to take advantage of the customer (or customers with usage patterns similar to mine at least).
Warp to Jupyter
In the course of my normal explorations and projects, I have been ‘working’ lately in a different software environment - Jupyter notebooks within VS Code. Warp does not seem to do as well with coding in this environment as it did with e.g. writing TypeScript/React code. A Jupyter notebook consists of a JSON-formatted .ipynb file that contains both embedded Python code and the output produced by this code (text, plots, etc). This mixture of code and output sections in the Jupyter .ipynb files greatly confuses Warp. I would have to remember to push a button in VS Code to remove all of the output from the Jupyter notebook before asking Warp to make any changes, else it would choke (and use lots of AI credits) trying to understand those outputs, especially the plots.
And, when Warp made changes to the Python code sections within a Jupyter .ipynb file, it would routinely corrupt the file by introducing JSON syntax errors. VS Code / Jupyter does not react well to having the .ipynb file externally changed such that the file becomes unparseable. After Warp corrupted the file, I would have to either find/fix the error manually using a different editor, or ask Warp to fix it, which it usually could do, at the cost of using more AI credits.
Unlike the pricing model changes discussed above, this reduction of the value proposition of Warp for me is not due to something that has changed in Warp. Instead my usage situation has changed to one where Warp doesn’t shine as brightly. I do think that Warp could be made better at dealing with Jupyter notebooks (by intrinsically understanding that it should ignore the output sections in an .ipynb file, or by making sure never to corrupt the JSON), but there may not be much customer demand to make Warp work better with Jupyter notebooks.
Charging for AI
The whole concept of a Warp AI credit seems a bit loose. AI platforms have traditionally priced usage based on tokens, a unit that approximates the number of words in a user’s input. One reason that Warp can do what it does is that it feeds in a large amount of code and prior interactions as context to each AI request. This could make even a simple AI request use a large number of tokens because of the code context and conversation history. So probably charging by the token was not realistic for Warp, so it had to invent something like AI credits. Maybe Warp’s AI credits are based on traditional AI tokens at some level, or maybe not.
From a Warp user perspective, different tasks take different numbers of AI credits. Warp does a good job of telling you how many credits are being used as you go, which would seem to be necessary and good. But reporting the AI credit usage practically after every request also has a psychological downside. There sometimes seemed to be a disconnect between the complexity/impressiveness of the task that I asked Warp to do and the number of AI credits it used to do it. Sometimes it would completely refactor and add new functionality to some code and get it all absolutely correct in one go, and report using only 5 credits. Other times it would do something not nearly so impressive and report using 50 credits. It was especially disconcerting to see many credits used when it corrupted a file, then more credits used when I asked it to repair the corruption.
I do understand that the whole area of how to fairly price AI usage is an open question in the industry, so I can’t call out Warp’s AI credit scheme as being worse than how other companies do it. But the observed inconsistency in the number of Warp AI credits used for different tasks can lead one to believe that maybe Warp AI credits are defined in a soft, flexible way. If you start to feel that you are not getting the value out of Warp that you have in the past, you may start to wonder if a ‘floating’ definition of an AI credit is a way to cover up charging more money for the same amount of AI processing.
Warp Product Direction
Warp has made recent changes to their UI that are not useful and even get in the way for me. I use Warp as a ‘smart console’ in conjunction with other tools like VS Code. I think Warp’s concept of integrating AI with a terminal application is a strong one, and it is what attracted me to Warp in the first place. But Warp seems to be trying to turn into a complete standalone development environment of its own, with multiple editor windows and git integration. This includes a git diff window that always seemed to be present when I didn’t want it - I was constantly closing it. I suppose these changes are wonderful for some users, but not for me. It makes me think that Warp is evolving away from what attracted me to it in the first place.
Parting Shot
I had an unpleasant experience ‘on my way out the door’ with Warp. After I had already canceled my Warp subscription, I had a significant number of AI credits that I planned to use before the subscription actually ended. Warp shows the date when your AI credits will expire on a screen within Warp, so a couple of days before that date I went to use those credits on a project. But I found my Warp account was already closed and those AI credits had vanished. I sent a tersely-worded email to Warp support complaining about my lost credits and the closure of my account before the end date. They did at least respond to my email, but only to explain that this happened because in their complicated renewal scheme, the subscription renewal date (and thus the cancellation date) can get out of sync with the AI credits expiration date. So my subscription ended a few days before the AI credits expiration date that Warp was showing me (Warp does not show the subscription renewal date on the same screen). The response from support acknowledged that these dates being out of sync can be frustrating to customers and that they were working on improving this, but they did not offer to reinstate the lost credits.
So good riddance, Warp. I will be exploring other AI tools soon, probably starting with GitHub Copilot.
